Abstract: Ethylene cracker generally utilizes petroleum products as raw materials, and turn the long-chain hydrocarbon molecules into a variety of short chain hydrocarbons like gaseous hydrocarbons and a small amount of liquid hydrocarbons at high temperature, thereby obtaining ethylene, propylene and other products. Establishing the yield model for main products is of great significance to advanced control, operation optimization of the pyrolysis furnace in theory and practice. Due to the semi-continuous and periodic characteristics of cracking furnace, establishing a yield model for each condition is time-consuming. Considering the temperature, residence time and pressure have common effect on product yields, so this article uses these similarities exist in the new and old processes, supplemented with transfer learning to accomplish quick modeling for different conditions. Compared to previous studies, this method can be able to fully exploit the information contained in the historical data from a small amount of new data. Meanwhile, building a model of ethylene in the process of cracking furnace, case studies demonstrate the efficacy of the developed methodology.

Key words: similarities, ethylene cracking furnace, transfer learning, yields, prediction, modeling, neural networks

摘要： 乙烯裂解炉通常以石油分馏产品为原料,并在高温条件下使长链分子的烃断裂成各种短链的气态烃和少量液态烃,从而获得乙烯、丙烯等烯烃及其他产品.建立这些主要产品的产率模型对裂解炉的先进控制、操作优化等任务在理论和实际上都具有重要意义.尽管在不同裂解原料、不同炉型的裂解炉状况下产品收率均存在差异,但由于裂解炉运行具有半连续性、周期性特征,裂解温度、停留时间及烃分压等因素对裂解产率的影响具有共性,因此为减小建模过程中典型样本采集等成本,有效利用历史数据提高建模精度,有效利用这些不同运行过程中存在的相似性,辅以迁移学习算法实现不同工况下裂解产率的快速建模.相比较以前的研究,此建模方法在少量新数据的情况下充分挖掘了历史数据中包含的信息.最后,以某乙烯厂为研究实例进行裂解产率建模,结果显示能够获得较好的产率预测精度,验证了该建模方法的有效性.

关键词: 相似性, 乙烯裂解炉, 迁移学习, 产率, 预测, 模型, 神经网络